pith. sign in

arxiv: 2605.18303 · v1 · pith:5WLJNG2Knew · submitted 2026-05-18 · 💻 cs.LG · cs.AI· cs.CV· cs.RO

PH-Dreamer: A Physics-Driven World Model via Port-Hamiltonian Generative Dynamics

Xueyu Luan , Chenwei Shi This is my paper

Pith reviewed 2026-05-20 12:23 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.CVcs.RO
keywords port-hamiltonian dynamicsworld modelslatent imaginationphysics-informed learningvisual controlenergy conservationactor-critic
0
0 comments X

The pith

Port-Hamiltonian energy routing in latent dynamics produces compact world models that respect conservation laws and improve control performance.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Standard recurrent world models generate latent sequences efficiently but often break physical rules such as energy conservation and dissipation. The paper replaces unstructured transitions with a Port-Hamiltonian structure that treats latent changes as action-driven energy flow and dissipation. It adds a separate kinematics-aware module that reads proprioceptive signals to compute the Hamiltonian and power balance, then uses those energy gradients inside an actor-critic loop to penalize high-energy actions. The resulting models show higher final returns, closer matches between imagined and real rewards, and measurable shrinkage in latent volume and control effort.

Core claim

The authors present a unified Port-Hamiltonian framework that embeds physical priors directly into recurrent latent transitions by modeling them as action-controlled energy routing governed by flow and dissipation; this is paired with a kinematics-aware energy world model that extracts the Hamiltonian and power balance from observations and with an energy-guided actor-critic that applies Lagrangian regularization to favor lower-energy policies. On visual control benchmarks the approach yields superior asymptotic returns, tighter lower-variance alignment between imagined and real rewards, and concrete reductions in latent phase-space volume, energy use, and jerk.

What carries the argument

The Port-Hamiltonian framework, which projects latent evolution onto a phase space whose dynamics are defined by action-controlled energy routing, flow, and dissipation.

If this is right

  • Higher asymptotic returns on visual control benchmarks.
  • Tighter, lower-variance alignment between imagined and real rewards.
  • Latent phase space volume reduced by 4.18-8.41 percent.
  • Energy consumption lowered by up to 7.80 percent.
  • Mean squared jerk lowered by up to 9.38 percent.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same energy-routing prior could be applied to non-visual state spaces such as tactile or audio streams to enforce consistency in other modalities.
  • Reduced latent volume may shorten the horizon needed for accurate long-term planning without extra training data.
  • Energy-guided regularization might produce policies that remain stable when transferred to real hardware with different dynamics.
  • The Hamiltonian estimation module could serve as a diagnostic tool to detect when a world model begins to drift from physical plausibility.

Load-bearing premise

Modeling projected latent evolution as action-controlled energy routing governed by flow and dissipation will automatically produce a more compact and physically structured phase space without new inconsistencies.

What would settle it

Run the same visual-control benchmarks and check whether the measured reductions in latent volume, energy consumption, and jerk disappear or whether imagined-reward variance increases while physical conservation violations remain in the generated trajectories.

Figures

Figures reproduced from arXiv: 2605.18303 by Chenwei Shi, Xueyu Luan.

Figure 1
Figure 1. Figure 1: Architecture of the Port Hamiltonian world model. (a) Implicit structural constraints regularizing the recurrent state space transitions within the RSSM backbone. (b) Internal configuration of the implicit Hamiltonian estimator for unsupervised latent geometry supervision. (c) Network architecture for explicit Hamiltonian estimation grounded in proprioceptive kinematic observations. (d) Action driven energ… view at source ↗
Figure 2
Figure 2. Figure 2: Latent phase space analysis. PH-RSSM yields a more compact projected PH phase space than the baseline. On a shared input trajectory, its projected latents remain more bounded, supporting latent stability gains without implying a closed loop boundedness guarantee [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Explicit energy alignment. Predicted Hamiltonians are compared with MuJoCo ground truth mechanical energy across six DMC tasks. Each panel reports current-state and post-action energy predictions, with an inset highlighting the tail segment. fluctuations rather than relying solely on visual pattern memorization, supporting its use for energy aware policy regularization. Evaluation Metrics. To quantify the … view at source ↗
read the original abstract

World models built on recurrent state space architectures enable efficient latent imagination, yet remain physically unstructured, producing dynamics that violate conservation and dissipative principles. We introduce a unified Port-Hamiltonian framework that remedies this through three synergistic mechanisms. First, we embed implicit physical priors into recurrent transitions by modeling projected latent evolution as action controlled energy routing governed by flow and dissipation, biasing the projected PH phase space toward a more compact and physically structured representation. Second, we develop a kinematics aware energy world model that estimates the Hamiltonian and power balance from proprioceptive observations, providing an explicit physical signal for thermodynamic reasoning. Third, leveraging these energy gradients, we establish an energy guided Actor-Critic that uses Lagrangian multipliers to regularize policy optimization toward lower energy and smoother control. Across visual control benchmarks, this paradigm not only attains superior asymptotic returns but also elevates internal simulator fidelity by establishing a tighter, lower variance alignment between imagined and real rewards, all while reducing latent phase space volume by 4.18-8.41%, energy consumption by up to 7.80%, and mean squared jerk by up to 9.38%.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces PH-Dreamer, a unified Port-Hamiltonian framework for physics-driven world models in visual control. It embeds implicit physical priors into recurrent state-space transitions by modeling projected latent evolution as action-controlled energy routing with flow and dissipation terms. It develops a kinematics-aware energy world model that estimates the Hamiltonian and power balance from proprioceptive observations. It further proposes an energy-guided Actor-Critic that employs Lagrangian multipliers to regularize policy optimization toward lower energy and smoother control. The paper reports superior asymptotic returns, tighter lower-variance alignment between imagined and real rewards, and quantitative reductions in latent phase-space volume (4.18–8.41 %), energy consumption (up to 7.80 %), and mean squared jerk (up to 9.38 %).

Significance. If the port-Hamiltonian energy-balance identity is verifiably preserved by the learned recurrent transitions, the work would provide a principled route to injecting conservation and dissipation structure into latent world models, potentially yielding more compact, physically plausible representations and smoother policies. The combination of an explicit energy estimator with gradient-based policy regularization is a concrete contribution that could improve internal simulator fidelity in model-based RL.

major comments (2)
  1. [Section 3] Recurrent transition model (Section 3): the central claim that projected latent evolution follows port-Hamiltonian dynamics requires an explicit derivation or numerical check that dH/dt equals input power minus dissipation for the learned model. The abstract and architecture description supply neither; without this verification the reported reductions in latent phase-space volume, energy consumption, and jerk cannot be attributed to the physical priors rather than incidental regularization.
  2. [Section 5] Experimental results (Section 5): the abstract states specific percentage reductions (4.18–8.41 % phase-space volume, up to 7.80 % energy, up to 9.38 % jerk) and performance gains but provides no error bars, statistical tests, baseline implementation details, or description of how energy gradients are computed and regularized. These omissions make it impossible to assess whether the gains are robust or post-hoc selected.
minor comments (2)
  1. [Abstract] Abstract: the claim of 'superior asymptotic returns' and 'elevated internal simulator fidelity' would benefit from naming the specific visual control benchmarks and the exact baseline methods used for comparison.
  2. [Section 4] Notation: clarify the precise functional form of the kinematics-aware energy estimator and how the power-balance term is obtained from proprioceptive observations alone.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thoughtful and constructive comments on our manuscript. We address each major comment below, providing clarifications where appropriate and outlining the specific revisions we will incorporate to strengthen the presentation and verifiability of our results.

read point-by-point responses

  1. Referee: [Section 3] Recurrent transition model (Section 3): the central claim that projected latent evolution follows port-Hamiltonian dynamics requires an explicit derivation or numerical check that dH/dt equals input power minus dissipation for the learned model. The abstract and architecture description supply neither; without this verification the reported reductions in latent phase-space volume, energy consumption, and jerk cannot be attributed to the physical priors rather than incidental regularization.

    Authors: The recurrent transition is formulated by construction to obey port-Hamiltonian dynamics: the latent state evolution is defined via the skew-symmetric interconnection matrix, the dissipation matrix, and the input port, which analytically guarantees that dH/dt equals the supplied power minus the dissipated power. We will add an explicit derivation of this energy-balance identity to Section 3 in the revision. We will also include a numerical verification subsection that evaluates the identity on the trained model across all benchmarks, reporting the residual error to confirm that the observed reductions in phase-space volume, energy, and jerk can be attributed to the enforced physical structure rather than generic regularization. revision: yes

  2. Referee: [Section 5] Experimental results (Section 5): the abstract states specific percentage reductions (4.18–8.41 % phase-space volume, up to 7.80 % energy, up to 9.38 % jerk) and performance gains but provides no error bars, statistical tests, baseline implementation details, or description of how energy gradients are computed and regularized. These omissions make it impossible to assess whether the gains are robust or post-hoc selected.

    Authors: We agree that the current experimental section lacks sufficient statistical rigor and implementation transparency. In the revised manuscript we will (i) report all metrics with mean and standard deviation over at least five independent random seeds, (ii) include paired t-tests or Wilcoxon tests with p-values to establish statistical significance, (iii) expand the baseline descriptions with exact hyper-parameter settings and code references, and (iv) provide a detailed derivation of the energy-gradient computation together with the precise form of the Lagrangian multiplier regularization used in the Actor-Critic. These additions will allow readers to assess robustness and rule out post-hoc selection. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation relies on architectural priors and empirical measurement

full rationale

The paper defines a port-Hamiltonian recurrent transition by modeling latent evolution as action-controlled energy routing with explicit flow and dissipation terms, then separately learns a kinematics-aware Hamiltonian estimator from proprioceptive observations and applies its gradients via Lagrangian-regularized actor-critic. The reported reductions in latent phase-space volume, energy consumption, and jerk are presented as measured outcomes on visual control benchmarks after training, not as quantities that are definitionally identical to the fitted parameters or enforced identities. No load-bearing self-citation, ansatz smuggling, or renaming of known results appears; the physical structure is an input modeling choice whose downstream effects are evaluated externally rather than tautologically recovered from the same fit.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only review prevents exhaustive extraction; the framework implicitly relies on the assumption that port-Hamiltonian structure can be projected onto latent recurrent transitions without violating the original physical conservation properties.

axioms (1)
  • domain assumption Port-Hamiltonian systems preserve energy balance through explicit flow and dissipation terms.
    Invoked when the paper states that projected latent evolution is governed by flow and dissipation.

pith-pipeline@v0.9.0 · 5734 in / 1382 out tokens · 29861 ms · 2026-05-20T12:23:53.450988+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

38 extracted references · 38 canonical work pages · 5 internal anchors

  1. [1]

    Christensen, Hao Su, Jiajun Wu, and Yunzhu Li

    Bo Ai, Stephen Tian, Haochen Shi, Yixuan Wang, Tobias Pfaff, Cheston Tan, Henrik I. Christensen, Hao Su, Jiajun Wu, and Yunzhu Li. A review of learning-based dynamics models for robotic manipulation. Science Robotics, 10 0 (106): 0 eadt1497, 2025

    work page 2025

  2. [2]

    On the analysis of movement smoothness

    Sivakumar Balasubramanian, Alejandro Melendez-Calderon, Agnes Roby-Brami, and Etienne Burdet. On the analysis of movement smoothness. Journal of NeuroEngineering and Rehabilitation, 12 0 (1): 0 112, 2015

    work page 2015

  3. [3]

    Kochmann

    Jan - Hendrik Bastek, WaiChing Sun, and Dennis M. Kochmann. Physics-informed diffusion models. In International Conference on Learning Representations (ICLR), 2025

    work page 2025

  4. [4]

    Video generation models as world simulators, 2024

    Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators, 2024. URL https://openai.com/index/video-generation-models-as-world-simulators/. OpenAI technical report

    work page 2024

  5. [5]

    Smith, Kelsey R

    Filipe de Avila Belbute - Peres, Kevin A. Smith, Kelsey R. Allen, Josh Tenenbaum, and J. Zico Kolter. End-to-end differentiable physics for learning and control. In Advances in Neural Information Processing Systems, volume 31, pages 7178--7189, 2018

    work page 2018

  6. [6]

    DreamerPro : Reconstruction-free model-based reinforcement learning with prototypical representations

    Fei Deng, Ingook Jang, and Sungjin Ahn. DreamerPro : Reconstruction-free model-based reinforcement learning with prototypical representations. In Proceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pages 4956--4975. PMLR, 2022

    work page 2022

  7. [7]

    T rans D reamer V 3: Implanting T ransformer in D reamer V 3

    Shruti Sadanand Dongare, Amun Kharel, Jonathan Samuel, and Xiaona Zhou. T rans D reamer V 3: Implanting T ransformer in D reamer V 3. arXiv preprint arXiv:2506.17103, 2025

    work page arXiv 2025

  8. [8]

    FOCUS : object-centric world models for robotic manipulation

    Stefano Ferraro, Pietro Mazzaglia, Tim Verbelen, and Bart Dhoedt. FOCUS : object-centric world models for robotic manipulation. Frontiers in Neurorobotics, 19: 0 1585386, 2025

    work page 2025

  9. [9]

    G enesis: A generative and universal physics engine for robotics and beyond, December 2024

    Genesis Authors . G enesis: A generative and universal physics engine for robotics and beyond, December 2024. URL https://github.com/Genesis-Embodied-AI/Genesis

    work page 2024

  10. [10]

    H amiltonian neural networks

    Samuel Greydanus, Misko Dzamba, and Jason Yosinski. H amiltonian neural networks. In Advances in Neural Information Processing Systems, volume 32, pages 15353--15363, 2019

    work page 2019

  11. [11]

    T 2 V P hys B ench: A first-principles benchmark for physical consistency in text-to-video generation

    Xuyang Guo, Jiayan Huo, Zhenmei Shi, Zhao Song, Jiahao Zhang, and Jiale Zhao. T 2 V P hys B ench: A first-principles benchmark for physical consistency in text-to-video generation. arXiv preprint arXiv:2505.00337, 2025

    work page arXiv 2025

  12. [12]

    Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee, and James Davidson

    Danijar Hafner, Timothy P. Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee, and James Davidson. Learning latent dynamics for planning from pixels. In Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 2555--2565. PMLR, 2019

    work page 2019

  13. [13]

    Lillicrap, Jimmy Ba, and Mohammad Norouzi

    Danijar Hafner, Timothy P. Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. In International Conference on Learning Representations (ICLR), 2020

    work page 2020

  14. [14]

    Lillicrap, Mohammad Norouzi, and Jimmy Ba

    Danijar Hafner, Timothy P. Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. In International Conference on Learning Representations (ICLR), 2021

    work page 2021

  15. [15]

    Lillicrap

    Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy P. Lillicrap. Mastering diverse control tasks through world models. Nature, 640 0 (8059): 0 647--653, 2025 a

    work page 2025

  16. [16]

    Training Agents Inside of Scalable World Models

    Danijar Hafner, Wilson Yan, and Timothy P. Lillicrap. Training agents inside of scalable world models. arXiv preprint arXiv:2509.24527, 2025 b

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  17. [17]

    S afe D reamer: Safe reinforcement learning with world models

    Weidong Huang, Jiaming Ji, Chunhe Xia, Borong Zhang, and Yaodong Yang. S afe D reamer: Safe reinforcement learning with world models. In International Conference on Learning Representations (ICLR), 2024

    work page 2024

  18. [18]

    How far is video generation from world model: A physical law perspective

    Bingyi Kang, Yang Yue, Rui Lu, Zhijie Lin, Yang Zhao, Kaixin Wang, Gao Huang, and Jiashi Feng. How far is video generation from world model: A physical law perspective. In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, pages 28991--29017. PMLR, 2025

    work page 2025

  19. [19]

    Learning to walk from three minutes of real-world data with semi-structured dynamics models

    Jacob Levy, Tyler Westenbroek, and David Fridovich - Keil. Learning to walk from three minutes of real-world data with semi-structured dynamics models. In Proceedings of The 8th Conference on Robot Learning, volume 270 of Proceedings of Machine Learning Research, pages 2061--2079. PMLR, 2025

    work page 2061

  20. [20]

    PIN-WM : Learning physics-informed world models for non-prehensile manipulation

    Wenxuan Li, Hang Zhao, Zhiyuan Yu, Yu Du, Qin Zou, Ruizhen Hu, and Kai Xu. PIN-WM : Learning physics-informed world models for non-prehensile manipulation. In Robotics: Science and Systems XXI, 2025

    work page 2025

  21. [21]

    P hys G en: Rigid-body physics-grounded image-to-video generation

    Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. P hys G en: Rigid-body physics-grounded image-to-video generation. In European Conference on Computer Vision (ECCV), volume 15140, pages 360--378, 2024

    work page 2024

  22. [22]

    Deep Lagrangian networks: Using physics as model prior for deep learning

    Michael Lutter, Christian Ritter, and Jan Peters. Deep Lagrangian networks: Using physics as model prior for deep learning. In International Conference on Learning Representations (ICLR). OpenReview.net, 2019

    work page 2019

  23. [23]

    R2-Dreamer : Redundancy-reduced world models without decoders or augmentation

    Naoki Morihira, Amal Nahar, Kartik Bharadwaj, Yasuhiro Kato, Akinobu Hayashi, and Tatsuya Harada. R2-Dreamer : Redundancy-reduced world models without decoders or augmentation. In International Conference on Learning Representations (ICLR), 2026

    work page 2026

  24. [24]

    SOLD : Slot object-centric latent dynamics models for relational manipulation learning from pixels

    Malte Mosbach, Jan Niklas Ewertz, Angel Villar - Corrales, and Sven Behnke. SOLD : Slot object-centric latent dynamics models for relational manipulation learning from pixels. In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, pages 44911--44935. PMLR, 2025

    work page 2025

  25. [25]

    Do generative video models understand physical principles?

    Saman Motamed, Laura Culp, Kevin Swersky, Priyank Jaini, and Robert Geirhos. Do generative video models understand physical principles? arXiv preprint arXiv:2501.09038, 2025

    work page internal anchor Pith review arXiv 2025

  26. [26]

    Representation Learning with Contrastive Predictive Coding

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  27. [27]

    Physics-informed model-based reinforcement learning

    Adithya Ramesh and Balaraman Ravindran. Physics-informed model-based reinforcement learning. In Proceedings of The 5th Annual Learning for Dynamics and Control Conference, volume 211 of Proceedings of Machine Learning Research, pages 26--37. PMLR, 2023

    work page 2023

  28. [28]

    Design principles for energy-efficient legged locomotion and implementation on the MIT cheetah robot

    Sangok Seok, Albert Wang, Meng Yee Chuah, Dong Jin Hyun, Jongwoo Lee, David M Otten, Jeffrey H Lang, and Sangbae Kim. Design principles for energy-efficient legged locomotion and implementation on the MIT cheetah robot. IEEE/ASME Transactions on Mechatronics, 20 0 (3): 0 1117--1129, 2015

    work page 2015

  29. [29]

    Roboscape: Physics-informed embodied world model

    Yu Shang, Xin Zhang, Yinzhou Tang, Lei Jin, Chen Gao, Wei Wu, and Yong Li. Roboscape: Physics-informed embodied world model. arXiv preprint arXiv:2506.23135, 2025

    work page arXiv 2025

  30. [30]

    Learning symbolic models for graph-structured physical mechanism

    Hongzhi Shi, Jingtao Ding, Yufan Cao, Quanming Yao, Li Liu, and Yong Li. Learning symbolic models for graph-structured physical mechanism. In International Conference on Learning Representations (ICLR), 2023

    work page 2023

  31. [31]

    Learning latent dynamic robust representations for world models

    Ruixiang Sun, Hongyu Zang, Xin Li, and Riashat Islam. Learning latent dynamic robust representations for world models. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 47234--47260. PMLR, 2024

    work page 2024

  32. [32]

    DeepMind Control Suite

    Yuval Tassa, Yotam Doron, Alistair Muldal, Tom Erez, Yazhe Li, Diego de Las Casas, David Budden, Abbas Abdolmaleki, Josh Merel, Andrew Lefrancq, Timothy P. Lillicrap, and Martin A. Riedmiller. D eep M ind control suite. arXiv preprint arXiv:1801.00690, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  33. [33]

    Rezende, Andrew Jaegle, S \' e bastien Racani \` e re, Aleksandar Botev, and Irina Higgins

    Peter Toth, Danilo J. Rezende, Andrew Jaegle, S \' e bastien Racani \` e re, Aleksandar Botev, and Irina Higgins. H amiltonian generative networks. In International Conference on Learning Representations (ICLR), 2020

    work page 2020

  34. [34]

    Learning physical constraints with neural projections

    Shuqi Yang, Xingzhe He, and Bo Zhu. Learning physical constraints with neural projections. In Advances in Neural Information Processing Systems, volume 33, 2020

    work page 2020

  35. [35]

    Fast and efficient locomotion via learned gait transitions

    Yuxiang Yang, Tingnan Zhang, Erwin Coumans, Jie Tan, and Byron Boots. Fast and efficient locomotion via learned gait transitions. In Proceedings of the 5th Conference on Robot Learning, volume 164 of Proceedings of Machine Learning Research, pages 773--783. PMLR, 2022

    work page 2022

  36. [36]

    Task aware dreamer for task generalization in reinforcement learning

    Chengyang Ying, Xinning Zhou, Zhongkai Hao, Hang Su, Songming Liu, Dong Yan, and Jun Zhu. Task aware dreamer for task generalization in reinforcement learning. arXiv preprint arXiv:2303.05092, 2023

    work page arXiv 2023

  37. [37]

    STORM : Efficient stochastic transformer based world models for reinforcement learning

    Weipu Zhang, Gang Wang, Jian Sun, Yetian Yuan, and Gao Huang. STORM : Efficient stochastic transformer based world models for reinforcement learning. In Advances in Neural Information Processing Systems, volume 36, pages 27147--27166, 2023

    work page 2023

  38. [38]

    VBench-2.0: Advancing Video Generation Benchmark Suite for Intrinsic Faithfulness

    Dian Zheng, Ziqi Huang, Hongbo Liu, Kai Zou, Yinan He, Fan Zhang, Yuanhan Zhang, Jingwen He, Wei - Shi Zheng, Yu Qiao, and Ziwei Liu. V B ench-2.0: Advancing video generation benchmark suite for intrinsic faithfulness. arXiv preprint arXiv:2503.21755, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025